researchsquare.comSource publication:
Loading content…
Generative enzyme design has mostly optimized at the amino acid level, rarely accounting for coding sequences and host expression at once. EZPro-Fit places protein 3D structure, coding sequences and taxonomic information in one generative framework, pretrained on thermophilic microbial data and fine-tuned on highly expressed E. coli genes, so it generates host-adapted coding sequences directly from structure. On the industrial endoglucanase Cel5E, its variants N291D and D259E raised both melting temperature and catalytic activity in experiments.
Enzyme variant design can now skip the two-step route of first designing an amino acid sequence and then optimizing codons, generating coding sequences expressible in a target host directly from a protein backbone. Earlier structure-to-sequence models output amino acid sequences only and lack DNA-level prediction; thermostability-oriented tools generally require a predefined wild-type sequence or an enumerated mutation set and search the sequence space locally. On a held-out test set of thermostable protein structures, the full framework reached 55.54% residue recovery, a 17.4% relative improvement over ProteinMPNN, and 32.34% CDS recovery, a 13.8% relative improvement over CodonMPNN, while keeping a TM-score of 0.89 and a mean RMSD of 3.16 Å.
The framework jointly models structural geometry, coding sequences and species-of-origin information, so generated coding sequences keep protein-level constraints while favoring host-preferred codons. Conventional practice decouples protein sequence design from downstream nucleotide optimization, and translational information is not explicitly represented in structure or amino acid representations. An MPNN encoder turns the protein backbone into latent vectors, conditional flow matching generates sequences, and a dual-decoder feedback loop scores amino acid sequences with a pretrained ESM2; pretraining used thermophilic bacteria with optimal growth temperature at or above 40°C, dereplicated at 40% sequence identity and filtered to mean pLDDT above 90, while fine-tuning used E. coli K-12 MG1655 genes with TPM above 500.
On Cel5E, variants proposed by the framework were tested in wet lab and included sites that improved thermostability and catalytic activity at the same time. Among the thirteen tested variants in the same batch, most substitutions improved only one property or had marginal effects, and the double mutant DS reached the highest melting temperature of 78.12°C but retained only about one third of wild-type activity. N291D raised the melting temperature from 63.16°C for the wild type to 69.58°C while increasing relative activity by about 35%; D259E raised the melting temperature to 72.35°C with a 17% activity improvement.
The demonstrated scope is protein structure data from thermophilic bacteria and a fine-tuning setup with E. coli as the expression host; experimental validation centers on thirteen variants of the industrial endoglucanase Cel5E. Other hosts or enzyme families would require preparing corresponding highly expressed gene data for host adaptation.
PSSM scores of designed sequences are lower than native sequences across all conservation bins, so reproduction of native evolutionary preferences remains incomplete; after fine-tuning the %MINMAX metric drops from about 79.2–81.8 to about 76.5–77.8, a slight change in local codon usage. The thermostability-associated amino acid preference term is explicitly described as a composition-level regularizer rather than a direct predictor of melting temperature or folding free energy.